Choosing a Reference Database

Getting reference choice right matters as much as getting the wet lab right, and it's one of the most commonly rushed decisions in the whole pipeline.

We've flagged this as one of the trickier parts of the workflow more than once elsewhere on this site (see Data Prep Short-Reads (PE), Step F), and promised a page that actually walks through it properly. This is that page. Taxonomic assignment quality depends on your classifier and its parameters, but it depends just as much, arguably more, on the reference you point it at. A great classifier pointed at the wrong reference will still give you confidently wrong answers.

Coverage and Fit Matter More Than Size

The single most important question is whether the reference actually covers the region your primers amplify. If a database is built from a different section of the gene, or only partially overlaps your amplicon, matching against it will be poor no matter how well-curated or comprehensive the database otherwise is. Fit comes before everything else.

Size is not the goal. A bigger reference isn't automatically better, and records unrelated to your actual study system, taxonomically "exotic" entries with no real chance of appearing in your samples, don't just fail to help, they actively make classification worse by giving the classifier more ways to be confidently wrong.

Public, general-purpose databases like SILVA (16S) or UNITE (ITS) are a convenient starting point, and often a reasonable one, but convenient isn't the same as correct for your specific project. Treat them as a default to evaluate against your actual amplicon and study system, not an automatic choice.

Reliability: A Few Mistakes Are Tolerable, Not Many

Reference records need to be accurate. Most classifiers, SINTAX included (see Step F for how its confidence thresholds work), tolerate some level of error in the reference without falling apart, mislabeled records, minor sequencing errors in the reference itself. But tolerance has a limit, and errors compound: a reference with too many inaccuracies doesn't fail gracefully, it fails quietly, producing assignments that look confident but aren't trustworthy.

Diversity Within Taxa: It's About Honesty, Not Depth

It's tempting to think the goal of a good reference is pushing assignments as deep as possible, ideally to species level. That's not quite right. What you actually want is a reference with enough within-taxon sequence diversity that the classifier's confidence values mean what they claim to mean.

This is the same principle behind the data loss philosophy elsewhere on this site: the goal isn't to maximise a number, it's to make sure the number you get is honest. A reference that's too sparse within a taxon can make an assignment look more confident than it should be, simply because there was nothing in the database close enough to expose the uncertainty. A well-populated reference will sometimes give you a less satisfying answer, genus rather than species, for example, but a more honest one.

Don't Merge Databases Casually

Combining reference databases from different sources to improve coverage is tempting, but it's only safe if their taxonomic nomenclature is fully harmonised. If one source calls something by a different name, a different rank structure, or a different synonym than another, merging them doesn't just create a messier reference, it introduces real classification errors: the same organism can end up split across two different names, or two different organisms can end up merged under one. This isn't a cosmetic inconsistency to clean up later, it actively corrupts the reference's data.

Structural Gaps Persist as Databases Grow

References keep growing, but growth doesn't fix everything. Taxa that are difficult or impossible to culture remain chronically under-represented, or entirely absent, no matter how large the overall database gets, because most sequences still ultimately trace back to cultured isolates or well-sampled environments. This is a structural gap in how reference databases are built, not something that resolves itself with time. If your study system includes taxa that are hard to culture, expect this gap to affect you specifically, and don't assume a newer, larger database release has necessarily closed it.

Always Include Outgroups

A reference restricted only to your target group is a specific, concrete failure mode, not just a theoretical concern. Consider a COI database built only from your target group, say, a specific insect order. Primers described as "specific" to that group still sometimes amplify off-target material, plastid DNA, bacterial DNA, or other organisms entirely, that happens to share enough sequence similarity to bind the primer sites. If your reference contains nothing but your target group, the classifier has no correct answer available for that off-target read. It doesn't get flagged as unclassifiable, it gets force-assigned somewhere inside your target group instead, because that's the only thing in the reference to assign it to. A reference with appropriate outgroups gives the classifier somewhere honest to put a read that doesn't actually belong to your target group.

Full-Length vs. Partial-Length References

Longer amplicons don't automatically classify better, partly because of the outgroup and diversity issues above, and partly because of how most reference databases were actually built. See the Reference Database Coverage section of the Long-Read vs. Short-Read page for the specific case of full-length 16S: most SILVA and GTDB records are partial sequences derived from short-read studies, so a longer, more informative query can end up matching less completely against what's actually available, the opposite of what you'd expect from more information.

Building a Custom Reference

When no existing database fits well enough, building a project-specific one is an option. See Step F for how we approach this, curating candidate sequences from BLAST searches or applying a lowest common ancestor (LCA) approach to build a SINTAX-compatible reference. This is real, sometimes substantial work, and it's usually project-specific rather than reusable, but it's the right answer when the coverage and fit problems above can't be solved any other way.

Choosing a Classifier

Reference and classifier choice are related but separate decisions. We use SINTAX by default, but it isn't the only option. See Insights & Solutions for a comparison with RDP Classifier and QIIME 2's classifier plugins, and for PR2 and MIDORI as additional database options beyond SILVA and UNITE.

Summary

Consideration	What to check
Coverage and fit	Does the reference actually span the region your primers amplify?
Size	Bigger isn't better; irrelevant records make classification worse, not better
Reliability	Is the reference well-curated, with an error rate low enough for your classifier to tolerate?
Within-taxon diversity	Enough representative sequences per taxon that confidence values are honest, not just deep
Merging sources	Only if nomenclature is fully harmonised between them
Structural coverage gaps	Does your study system include hard-to-culture taxa likely under-represented regardless of database size?
Outgroups	Does the reference include plausible off-target lineages, not just your focal group?
Amplicon length	Longer isn't automatically better classified; check actual reference coverage at your amplicon length
Custom reference	Worth building when no existing database fits well enough, expect real effort